Block 1 / Topic 1.2 — Prompt level set + AI-as-Coach HO

1.2.1

The naked prompt problem

3 min Lecture

Slide 1 / 2 · The naked prompt

This is a naked prompt

Summarize this codebase.

6 words. Zero scaffolding. Reliable garbage.

Slide 2 / 2 · Three runs

Same prompt, three runs

Run 1: bullet list, no structure
Run 2: prose paragraph, different framing
Run 3: half-correct facts, no schema

Non-deterministic by accident, not by design.

Speaker notes

Slide 1 — name the pattern: "This is what most developers write. Six words. 'Summarize this codebase.' It's a naked prompt — no scaffolding, no role, no examples, no structure. And here's the thing: it works just well enough to make you think prompts are unreliable. They're not. Naked prompts are unreliable."

Slide 2 — the three runs: "Same naked prompt, three runs. Different shape every time. Different facts. Different framing. Most teams ship something like this, see the variance, and call it 'AI being non-deterministic.' That's not non-determinism — that's underspecification. We're going to fix it by adding the five components — role, examples, CoT, constraints, safety — and you'll direct an AI coach through it."

Anchor: "By the end of this topic, you'll have produced a v5 prompt with an AI coach — the 5 components as your rubric. That prompt is what you feed into the skill-creator in Block 2."

1.2.2

The six prompt components

3 min Lecture

Slide 1 / 2 · Components

The anatomy of a real prompt

#	Component	What it does
1	Role	Who the model is supposed to be
2	Task	What it's supposed to do
3	Context	What it needs to know
4	Examples	What "good" looks like
5	Format	The output's shape
6	Constraints	The rules it must follow

Slide 2 / 2 · Anatomy

Naked prompts skip 5 of 6

"Summarize this codebase" has only task
Every layer we add is one of these components
Most prompt failures are missing-component failures, not "model is dumb"

1.2.3

The ambiguity tax

3 min Lecture

Slide 1 / 2 · Clarity

The Ambiguity Tax

Every ambiguous word in your prompt = a coin flip
The model resolves ambiguity by guessing
More ambiguous words = more guesses = more variance
Clarity is the single highest-leverage edit

Slide 2 / 2 · Before / after

Before / after

Before:

Summarize this codebase.

After:

Produce a one-page briefing of this codebase
covering: architecture, entry points, data flow,
key dependencies, test patterns.

"Summarize" was the tax. We just paid it down.

1.2.4

Chain-of-thought

4 min Lecture

Slide 1 / 3 · CoT

The highest-leverage technique

Tell the model to think in steps before answering
Doubles or triples accuracy on multi-step tasks
Two flavors: explicit steps, or "let's think step by step"
Free. No new tooling. Just words.

Slide 2 / 3 · Applied

CoT applied to repo inspection

First, scan the directory structure.
Then, identify the entry points.
Then, trace data flow from entry to output.
Then, summarize.

4 steps. Each constrains the next.

Slide 3 / 3 · When to use

Use CoT when…

The task has > 1 logical step
Errors compound across steps
You want to audit the model's reasoning
Output should reflect a process, not a guess

Speaker notes

Slide 1: "Chain-of-thought is the highest-leverage prompt technique that exists. Just tell the model to think in steps before answering. Doubles or triples accuracy on multi-step tasks. It's free — no new tooling, no fine-tuning, just words."

Slide 2 — show it applied: "This is CoT for Project Inspector. Four named steps. Each step constrains the next. The model can't skip ahead and hallucinate. It has to scan first, identify entry points second, trace flow third, then summarize."

Slide 3 — when to reach for it: "CoT pays the most when the task has multiple steps, when errors compound, when you want to audit the reasoning. For repo inspection? All three. So CoT is non-negotiable — when your AI coach proposes a v5 without numbered steps, push back."

Anchor: "Atom: Atom - Chain-of-Thought (CoT) Prompting. Single biggest leverage point in this whole block."

1.2.5

★ The 5 components — and when to use each

4 min Lecture LOAD-BEARING — DO NOT CUT

AI will write your prompts. Knowing what makes up a prompt — and when each piece earns its keep — is the durable skill.

Slide 1 / 3 · The components

Five components. Five "when" signals.

Layer	Component	When you need it
0 — Naked	Just the task	Single-shot, throwaway, exploration
1 — Role	Who the model is	Output needs perspective or domain framing
2 — Examples	One worked example	Output shape is non-obvious or has many valid forms
3 — CoT	Steps before answer	Task has multiple logical steps; errors compound
4 — Constraints	Format, schema, output shape	Output will be parsed, validated, or fed downstream
5 — Safety	What not to do	Output goes to production, users, or regulated context

Slide 2 / 3 · The durable skill

AI writes. You direct.

You will not hand-author prompts in production. You'll work with AI to author them.
But the AI doesn't know which components your task needs.
Your judgment fills that gap: "add an example here," "this needs a schema," "this is going to prod — add safety."
The 5 components are the vocabulary you bring to that conversation.
Knowing when matters more than knowing how to type.

Slide 3 / 3 · Next

Now you direct an AI coach

Pick any AI chat — Claude, Gemini, ChatGPT
Paste our optimizing prompt + your starting seed
The AI asks you questions about your task; you bring the judgment
Together you produce a v5 prompt for Project Inspector
The v5 you produce feeds Block 2 as your skill body

Speaker notes

Slide 1 — flip the table reading: "Five components. The same five we've been building up. But look at the third column — that's the column that matters when AI is writing your prompts. Not 'how to add a role,' but when do you need a role? Output needs perspective or domain framing. When do you need examples? When the shape is non-obvious. When does safety belong? When output goes to production."

Walk one or two rows out loud: "Look at row 4 — Constraints. When do you need them? When the output will be parsed or validated. If your prompt is producing JSON that another system reads, you need constraints. If it's producing prose for a human, maybe not. That's the judgment."

Slide 2 — the durable skill (memorize): "You will not hand-author prompts in production. You'll work with AI to author them. But the AI doesn't know which components your task needs. Your judgment fills that gap. The 5 components are the vocabulary you bring to that conversation. Knowing when matters more than knowing how to type."

Slide 3 — hand off to HO: "In the next 15 minutes, you direct an AI coach. Pick any chat tool — Claude, Gemini, ChatGPT. Paste our optimizing prompt and your starting seed. The AI asks; you bring the judgment about which components your task needs. The v5 prompt you produce is what feeds Block 2 — your skill body."

1.2.6

Advanced patterns awareness pass

3 min Lecture

Slide 1 / 2 · Three patterns

Beyond CoT — what else exists

Pattern	One-line use case
Tree of Thought (ToT)	Branch multiple reasoning paths; pick the best one
Self-Consistency	Run CoT N times; vote on the answer
ReAct	Interleave reasoning with tool calls (reason → act → observe → repeat)

Slide 2 / 2 · When they matter

When the 5 components aren't enough

ToT — when there are multiple plausible approaches
Self-Consistency — when you need confidence, not just an answer
ReAct — when the model needs to act on the world, not just think
Sam's block (Block 3) lives at the ReAct altitude in practice

Speaker notes

"Three more patterns you should know exist before Sam's block — we won't dive in, but you should be able to recognize them when they come up."

Walk the table fast — 30 seconds each:

"Tree of Thought — branch multiple reasoning paths, pick the best. Use when there are multiple plausible approaches and you want the model to compare."
"Self-Consistency — run CoT N times, vote on the answer. Use when you need confidence, not just an answer."
"ReAct — interleave reasoning with tool calls. Reason, act, observe, repeat. This is the altitude Sam's block lives at in practice — every Claude Code session is a ReAct loop under the hood."

Anchor: "Atoms on all three are in the vault if you want to read more. For now: knowing they exist is enough. You'll meet them when you need them."

1.2.7

★ HO — AI-as-Coach: collaborate to v5

15 min Hands-on LOAD-BEARING — V5 FEEDS BLOCK 2

You direct an AI coach to enhance your Project Inspector prompt. The 5 components are your vocabulary; the AI does the typing.

Slide 1 / 4 · The exercise

What you'll do (15 min)

Open any AI chat — Claude, Gemini, ChatGPT, your tool of choice
Copy the optimizing prompt (slide 2) → paste → send
The AI will ask for your prompt → reply with your starting seed (slide 3)
Answer the AI's 2–3 clarifying questions; bring judgment about which components your task needs
When the AI proposes a v2.0, check it against the rubric (slide 4)
Save the v5 — you'll paste it into a SKILL.md in Block 2

Working solo or with your neighbor — your call.

Slide 2 / 4 · The optimizing prompt

Copy this. Paste into your AI chat. Send.

Act as a Senior Prompt Engineer. Your goal is to help me refine and optimize
a prompt for maximum performance. When I provide a prompt, please follow these steps:

1. Critique: Review the prompt for structure, coherence, and clarity.
   Identify any ambiguous phrasing or "fluff" that might confuse an LLM.
2. Intent & Specificity: Evaluate if the prompt provides enough context
   and clear constraints. Identify what might be missing to achieve a high-quality result.
3. Clarifying Questions: Before providing the final version, ask me 2–3
   targeted questions to bridge any gaps in context, tone, or intended output format.
4. The Optimization: Once I answer, provide the "v2.0" version of the prompt
   using best practices (e.g., Role-assigning, Delimiters, Chain-of-Thought prompting).

Do you understand? If so, please ask me for the prompt you'd like to optimize.

[your current prompt]

Same optimizing prompt works in any AI chat. Pick whichever you brought. The [your current prompt] line at the end is a placeholder — your seed goes there in your next reply.

Slide 3 / 4 · Your follow-up reply

After the AI asks — send your seed

Minimal (Layer 0):

Summarize this codebase.

Slightly enriched (Layer 0+): if you want to start with a little context.

Summarize this codebase so a new engineer onboarding to the project can get oriented quickly.

Both are valid. Minimal lets the AI ask more clarifying questions; enriched starts the conversation further along. Your call.

Slide 4 / 4 · Self-check rubric

Your v5 should have…

Component	Should include
✓ ROLE	Senior engineer onboarding to a new project (or similar perspective)
✓ EXAMPLE	One worked briefing — shape of the output anchored
✓ STEPS	Numbered steps: scan → identify entry points → trace data flow → summarize
✓ FORMAT	JSON with named keys: `{architecture, entry_points, data_flow, key_dependencies, test_patterns}`
✓ SAFETY	No path hallucination; "unclear" when uncertain

If you hit all five, save it — that's your v5. Block 2 starts from there.

Speaker notes

Slide 1 — set the exercise: "Fifteen minutes. Any AI chat — Claude, Gemini, ChatGPT, whichever you brought. Copy our optimizing prompt. Add your seed. The AI will coach you. Your job is to bring judgment about which of the five components your task needs. By the end, you should have a v5 — and you check it against the rubric on slide 4."

Slide 2 — the copy: "This is the optimizing prompt. Copy the entire block. Paste it into your AI chat. The AI will read it and start playing the coach role — asking you about your task, suggesting components, proposing prompt text."

Slide 3 — starting seeds: "You'll need to give the AI a starting prompt to optimize. Two flavors — minimal (Layer 0, naked) or slightly enriched. Either is fine. Minimal forces the AI to ask more clarifying questions. Enriched lets you skip ahead. Your call based on how much guidance you want."

Slide 4 — the rubric: "Here's how you know you're done. Five boxes. If your v5 has all five, you're ready for Block 2. If you're missing one — go back and ask the AI to add it. The vocabulary you bring to that conversation is what makes this work."

Pace the room: set a 15-min timer. Walk the room every 3 min. Watch for two failure modes:

"My AI isn't asking me anything" — they probably pasted the seed without the optimizing prompt. Verify they pasted the optimizing prompt FIRST.
"My AI keeps going" — they need to use the rubric to recognize when they're done. Cue them to slide 4.

Stragglers (last 2 min): anyone without a v5 by 9:45 grabs the canonical fallback at prompts/project-inspector-v5.txt in the workshop repo. Don't surface this proactively — it undermines the exercise. Hand out 1:1 if needed.

Demo piece — instructor mirror

Instructor mirror: open Claude (or your AI of choice) on the projector. Paste the optimizing prompt + Layer 0 seed. Let the AI ask its first question. Answer it on screen so attendees see the rhythm.

Don't run the full conversation — just one or two exchanges, enough to model the pattern. Then say "now you go" and start the timer.

Optional — instructor reference v5 (do not show by default — undermines the exercise):

You are a senior engineer onboarding to a new project.
Your task: produce a structured briefing of the codebase.

Example briefing (small known project):
{
  "architecture": "Nuxt 3 SSR with Vue 3 composition API",
  "entry_points": ["app.vue", "pages/index.vue", "server/api/"],
  ...
}

Steps:
1. Scan the directory structure.
2. Identify entry points (look for app.vue, main.ts, index.*, package.json scripts).
3. Trace data flow from entry to output.
4. Summarize.

Output as JSON with exactly these keys:
{architecture, entry_points, data_flow, key_dependencies, test_patterns}

Safety:
- Never hallucinate file paths. Only describe what is observable in the provided files.
- If something is unclear, say "unclear" in the relevant field — do not guess.

Save target: attendees save as ~/code/ai-engineering-workshop/prompts/my-project-inspector-v5.txt (or anywhere they can find it again in Block 2).